SpaceGhost commited on
Commit
4094c87
·
verified ·
1 Parent(s): a311850

release(v12): official merged 16-bit safetensors release of SpaceCode-E2B V12

Browse files
MERGED_V12_MANIFEST.json ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": "v12",
3
+ "base_model": "google/gemma-4-E2B-it",
4
+ "adapter_source": "SpaceCode-E2B V12 Checkpoint 345 (9e250264e93278b65c14d98e27c9631fe293734a04ed7e27140f417fb1396361)",
5
+ "timestamp": "2026-11-01 00:16:35",
6
+ "total_size_bytes": 10242069243,
7
+ "files": {
8
+ "chat_template.jinja": {
9
+ "size_bytes": 18569,
10
+ "sha256": "0a2c8073c878ab1da004bee933a998606537bbb62016310352c7285c3f01c5b5"
11
+ },
12
+ "config.json": {
13
+ "size_bytes": 4644,
14
+ "sha256": "d9b2c908acf92ad058febae730771e1e3c85af051a63bf751dacb5b433ceee6c"
15
+ },
16
+ "generation_config.json": {
17
+ "size_bytes": 204,
18
+ "sha256": "2a143a0b852e68ee1ae1880f4ab4e4afea2bb3fcffb68a3cee4ef547182bc2e0"
19
+ },
20
+ "model-00001-of-00003.safetensors": {
21
+ "size_bytes": 1422278032,
22
+ "sha256": "67a2f4a7761f373975f639a83a97b6044353f4c682f8605776582377adfe55ad"
23
+ },
24
+ "model-00002-of-00003.safetensors": {
25
+ "size_bytes": 4987173642,
26
+ "sha256": "882030dc73cb8418a55a7e0b58ab6c81677330c19bb94518ea0a770542798057"
27
+ },
28
+ "model-00003-of-00003.safetensors": {
29
+ "size_bytes": 3800219068,
30
+ "sha256": "ec074aac9b6bca91363560f8e719c7476491b7db8660ec9a3535aea261d840d0"
31
+ },
32
+ "model.safetensors.index.json": {
33
+ "size_bytes": 200040,
34
+ "sha256": "4845b88fe98a87e97d3217b4c65c722951227eb6904a353022ae23f4fec3ebbb"
35
+ },
36
+ "processor_config.json": {
37
+ "size_bytes": 1689,
38
+ "sha256": "32bdf45d2ad4cc29a0822ddd157a182de76644f0419a6228d151495256e9813c"
39
+ },
40
+ "tokenizer.json": {
41
+ "size_bytes": 32169626,
42
+ "sha256": "cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f"
43
+ },
44
+ "tokenizer_config.json": {
45
+ "size_bytes": 3729,
46
+ "sha256": "3ab5c7b94dc97d65ca7064496fa69b88ff875378e1cb7ee3e43070c3a8170999"
47
+ }
48
+ }
49
+ }
Modelfile CHANGED
@@ -1,13 +1,14 @@
1
- FROM ./Q4_K_M.gguf
2
- MMPROJ ./mmproj-F16.gguf
3
 
4
- SYSTEM """Você é o SpaceCode-E2B, um assistente tutor especialista em Engenharia de Software e Desenvolvimento Mobile Android com React Native.
5
- Responda sempre com clareza, precisão técnica e código moderno seguindo as boas práticas da disciplina."""
 
 
6
 
7
- PARAMETER temperature 0.3
8
- PARAMETER top_p 0.95
9
- PARAMETER repeat_penalty 1.15
10
- PARAMETER stop "<|turn>"
11
- PARAMETER stop "<turn|>"
12
- PARAMETER stop "<|turn>user"
13
- PARAMETER stop "<bos>"
 
1
+ # Modelfile oficial para SpaceCode-E2B V12 (Ollama)
2
+ FROM ./SpaceCode-E2B-V12-Q4_K_M.gguf
3
 
4
+ PARAMETER temperature 0.0
5
+ PARAMETER stop "<end_of_turn>"
6
+ PARAMETER stop "<eos>"
7
+ PARAMETER stop "<start_of_turn>"
8
 
9
+ TEMPLATE """<start_of_turn>user
10
+ {{ .Prompt }}<end_of_turn>
11
+ <start_of_turn>model
12
+ """
13
+
14
+ SYSTEM """Você é o SpaceCode-E2B, assistente especializado em Engenharia de Software, Desenvolvimento Web/Mobile e Tutoria Técnica."""
 
README.md CHANGED
@@ -1,135 +1,201 @@
1
  ---
2
- license: apache-2.0
 
 
3
  base_model:
4
  - google/gemma-4-E2B-it
5
  tags:
6
  - code
7
  - vision
8
  - multimodal
9
- - gguf
10
- - llama.cpp
11
- - ollama
12
- - unsloth
13
  - software-engineering
14
  - react-native
15
  - android
16
- - expo
17
  - typescript
18
  - python
 
19
  language:
20
  - pt
21
  - en
22
  pipeline_tag: image-text-to-text
23
  ---
24
 
25
- # 🚀 SpaceCode-E2B (v2 — Release Oficial Homologado)
26
 
27
- **SpaceCode-E2B** é um modelo compacto de Engenharia de Software e Tutoria Educacional baseado no **Google Gemma 4 E2B (`google/gemma-4-E2B-it`)**, ajustado via QLoRA SFT com [Unsloth](https://github.com/unslothai/unsloth) pela **SpaceGhost**.
28
 
29
- > **Status de Homologação**: ✅ **Aprovado para Produção Educacional (v2 Merged 16-bit)**
30
- > **Benchmark Oficial (50 Questões Ementa Android/React Native)**: **44/50 (88.0%)** | **Score Composto**: **95.20%**
 
 
 
31
 
32
  ---
33
 
34
- ## 📊 Métricas Oficiais do Benchmark (50 Questões)
35
 
36
- Avaliação controlada sobre o currículo oficial de Desenvolvimento Mobile Android e React Native:
37
 
38
- | Tema da Ementa | Tópicos Cobertos | Acurácia v2 Oficial | Status |
39
- | :--- | :--- | :---: | :---: |
40
- | **Tema 2** | Sintaxe JSX, Expo CLI vs Bare CLI, NPM/Yarn, Hooks básicos | **6/8 (75.0%)** | Aprovado |
41
- | **Tema 3** | UI Nativa, FlatList, Flexbox, React Navigation, styled-components | **6/8 (75.0%)** | Aprovado |
42
- | **Tema 4** | Persistência Local (AsyncStorage, SQLite, Realm, MongoDB) | **8/8 (100.0%)** | **Perfeito** |
43
- | **Tema 5** | Conexão Remota (Fetch, Axios, OAuth2, Offline-First) | **7/8 (87.5%)** | Aprovado |
44
- | **Tema 6** | Arquitetura MVC/DAO, Flux vs Redux, AES/CryptoJS, Jest, Google Play | **7/8 (87.5%)** | Aprovado |
45
- | **Tema 7** | Desafios Integrados e Casos de Arquitetura | **10/10 (100.0%)** | **Perfeito** |
46
- | **TOTAL GERAL** | **50 Questões Fechadas de Múltipla Escolha** | **44/50 (88.0%)** | **Homologado** |
47
 
48
  ---
49
 
50
- ## 📦 Estrutura e Governança de Artefatos
51
 
52
- | Caminho no Repositório | Formato | Status | Descrição |
53
- | :--- | :--- | :---: | :--- |
54
- | `model.safetensors` (raiz) | **Safetensors (16-bit Merged)** | **Produção Oficial** | Modelo completo fundido (`v2 Merged`) para `transformers`, `vLLM` ou TGI (Texto + Visão + Áudio) |
55
- | `Q4_K_M.gguf` (raiz) | **GGUF (`Q4_K_M`)** | **Produção Oficial** | Pesos de linguagem quantizados em 4-bit para **Ollama** e **`llama.cpp`** |
56
- | `mmproj-F16.gguf` (raiz) | **GGUF Vision Projector (F16)** | **Produção Oficial** | Projetor multimodal (SigLIP/ViT) para leitura de imagens no **Ollama** e **`llama.cpp`** |
57
- | `lora/` | **QLoRA Adapter v2** | **Produção Oficial** | Pesos do adaptador validado com ~180 passos (`SpaceGhost/SpaceCode-E2B/lora`) |
58
- | `lora_v4/` | **Micro-LoRA Corretivo** | **Ativo Experimental** | Adaptador corretivo com 82.0% de acurácia global e correção do Caso 3 (`styled-components`) |
59
- | `lora_v3/` | **LoRA Pedagógico** | **Linha de Pesquisa** | Adaptador com estrutura explicativa detalhada de distratores de prova |
60
- | `lora_master/` | **Checkpoint Unificado** | **Arquivado** | Baseline em estudo de expansão de dados heterogêneos |
61
 
62
  ---
63
 
64
- ## 🚀 Como Executar
65
 
66
- ### 1. Via Ollama (Multimodal com Imagens + Texto)
67
- O repositório já inclui o `Modelfile` pré-configurado com ligação nativa ao `mmproj-F16.gguf`:
68
- ```bash
69
- # Baixar arquivos e registrar o modelo no Ollama:
70
- ollama create spacecode -f Modelfile
71
 
72
- # Executar com prompt de texto ou análise de imagem:
73
- ollama run spacecode "Analise esta tela de UI: /caminho/para/imagem.png"
74
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
75
 
76
- ### 2. Via `llama.cpp` (Texto + Visão)
 
 
 
 
77
  ```bash
78
- # Inferência multimodal com suporte a imagem:
79
- llama-cli -m Q4_K_M.gguf --mmproj mmproj-F16.gguf --image /caminho/para/imagem.png -p "<|turn>user\nExplique o problema nesta UI e mostre o código React Native correto.<turn|>\n<|turn>model\n"
80
  ```
81
 
82
- ### 3. Via Transformers (Multimodal Python Oficial)
83
  ```python
84
- from transformers import AutoProcessor, AutoModelForMultimodalLM
 
 
85
 
86
  MODEL_ID = "SpaceGhost/SpaceCode-E2B"
87
 
88
- # Carregar processador e modelo multimodal completo
89
  processor = AutoProcessor.from_pretrained(MODEL_ID)
90
- model = AutoModelForMultimodalLM.from_pretrained(
91
  MODEL_ID,
92
- torch_dtype="auto",
93
  device_map="auto"
94
  )
95
 
96
- # Processar texto e imagem:
97
- messages = [
98
  {
99
  "role": "user",
100
- "content": [
101
- {"type": "image", "url": "https://example.com/screenshot.png"},
102
- {"type": "text", "text": "Analise a tela e crie o componente React Native correspondente."}
103
- ]
104
  }
105
  ]
106
 
107
- inputs = processor.apply_chat_template(
108
- messages,
109
- tokenize=True,
110
- return_dict=True,
111
- return_tensors="pt",
112
- add_generation_prompt=True,
113
- ).to(model.device)
114
 
115
- outputs = model.generate(**inputs, max_new_tokens=512)
116
- response = processor.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
117
  print(response)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
118
  ```
119
 
 
 
120
  ---
121
 
122
- ## 🎯 Capacidades Homologadas
123
 
124
- 1. **Desenvolvimento Mobile Android (React Native & Expo)**:
125
- - Resolução de dúvidas conceituais e arquiteturais cobrindo Temas 2 ao 7.
126
- - Diagnóstico estático de código, verificação de dependências de hooks (`useEffect`), importação correta de módulos nativos e contratos assíncronos.
127
- 2. **Engenharia de Software e Reparo Causal Multimodal**:
128
- - Diagnóstico de falhas de renderização de interface visual (C08).
129
- - Geração de patches em *Unified Diff* e código corretivo com restauração visual comprovada.
130
 
131
  ---
132
 
133
- ## ⚠️ Atribuição
 
 
134
 
135
- - *SpaceCode-E2B is a fine-tuned derivative of Google Gemma 4 E2B (`google/gemma-4-E2B-it`).*
 
 
1
  ---
2
+ license: other
3
+ license_name: gemma-terms-of-use
4
+ license_link: https://ai.google.dev/gemma/terms
5
  base_model:
6
  - google/gemma-4-E2B-it
7
  tags:
8
  - code
9
  - vision
10
  - multimodal
11
+ - gemma4
 
 
 
12
  - software-engineering
13
  - react-native
14
  - android
 
15
  - typescript
16
  - python
17
+ - unsloth
18
  language:
19
  - pt
20
  - en
21
  pipeline_tag: image-text-to-text
22
  ---
23
 
24
+ # 🚀 SpaceCode-E2B (Versão Oficial V12 — Merged 16-bit)
25
 
26
+ **SpaceCode-E2B** é um modelo compacto e multimodal de alta precisão especializado em **Engenharia de Software, Desenvolvimento Mobile/Web (React Native, TypeScript, Python 3.12, Android) e Tutoria Acadêmica**, construído sobre a arquitetura **Google Gemma 4 E2B (`google/gemma-4-E2B-it`)** pela **SpaceGhost**.
27
 
28
+ > **Versão Oficial Atual:** `V12` (Release Tag: `v12`)
29
+ > **Formato de Distribuição:** Modelo completo fundido (*Merged 16-bit bfloat16 Safetensors*)
30
+ > **Modelo-Base:** `google/gemma-4-E2B-it` (Gemma4ForConditionalGeneration)
31
+ > **Adapter Homologado:** SpaceCode-E2B V12 Checkpoint 345 (`9e250264e93278b65c14d98e27c9631fe293734a04ed7e27140f417fb1396361`)
32
+ > **Status de Homologação:** ✅ **APROVADO E HOMOLOGADO PELO SOFTWARE ARCHITECT**
33
 
34
  ---
35
 
36
+ ## 📌 1. Visão Geral e Arquitetura
37
 
38
+ O SpaceCode-E2B V12 unifica as capacidades multimodais de visão e raciocínio técnico do Gemma 4 E2B com especialização rigorosa em código fortemente tipado, padrões arquiteturais modernos, tutoria socrática e mitigação sistemática de alucinações técnicas.
39
 
40
+ - **Arquitetura Base:** `Gemma4ForConditionalGeneration` (2.3B parâmetros de linguagem + Vision Tower SigLIP).
41
+ - **Precisão Numérica dos Pesos:** Bfloat16 (16-bit) completo, sem quantização destrutiva na exportação merged.
42
+ - **Tamanho Total em Disco:** 9,54 GiB (10.242.069.243 bytes distribuídos em 3 shards Safetensors).
43
+ - **Janela de Contexto Suportada:** 1.536 tokens nominais certificados no treinamento (expansível até 16.384 tokens pelo mecanismo de atenção nativo do Gemma 4).
 
 
 
 
 
44
 
45
  ---
46
 
47
+ ## 🛠️ 2. Informações do Treinamento e Fine-Tuning
48
 
49
+ - **Corpus de Treinamento:** V12 Sequence Gate (1.379 amostras ativas selecionadas com mascaramento estrito de prompt de sistema e labels pedagógicos).
50
+ - **Estratégia de Adaptação:** QLoRA SFT de alta fidelidade via Unsloth.
51
+ - Rank ($r$): 16
52
+ - Alpha ($\alpha$): 16
53
+ - Dropout: 0.05
54
+ - Target Modules: Camadas de projeção de atenção e MLP do modelo de linguagem (`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`).
55
+ - **Duração do Treinamento:** Exatamente 1 época experimental (345 passos do otimizador AdamW, com gradiente acumulado em 4 micro-lotes e último passo normalizado estritamente por 3.0 amostras).
56
+ - **Curva de Taxa de Aprendizado:** Cosine schedule com 10 passos de warmup ($2 \times 10^{-4} \to 2 \times 10^{-5}$).
57
+ - **Progressão de Perda (Loss):** Redução de 1,0356 (Passo 1) para 0,3568 (Passo 345), correspondendo a -65,55% de otimização.
58
 
59
  ---
60
 
61
+ ## 📊 3. Resultados da Avaliação Multimodal MMMU (TASK-0016)
62
 
63
+ Avaliação pareada e determinística (`temperature = 0.0`, `do_sample = False`) realizada sobre **100 questões oficiais do split de validação do benchmark MMMU (standard)**, distribuídas equitativamente entre 10 disciplinas acadêmicas (10 questões por disciplina):
 
 
 
 
64
 
65
+ | Disciplina Acadêmica | Grande Domínio | Gemma 4 E2B Original | SpaceCode-E2B V12 | Delta (p.p.) | Classificação |
66
+ | :--- | :--- | :---: | :---: | :---: | :--- |
67
+ | **Computer Science** | Tech & Engineering | 2/10 (20.0%) | **3/10 (30.0%)** | **+10.0%** | **SUPERIOR** |
68
+ | **Electronics** | Tech & Engineering | 0/10 (0.0%) | **3/10 (30.0%)** | **+30.0%** | **SUPERIOR** |
69
+ | **Mechanical Engineering** | Tech & Engineering | 0/10 (0.0%) | **1/10 (10.0%)** | **+10.0%** | **SUPERIOR** |
70
+ | **Math** | Science | 1/10 (10.0%) | **3/10 (30.0%)** | **+20.0%** | **SUPERIOR** |
71
+ | **Biology** | Science | 2/10 (20.0%) | **4/10 (40.0%)** | **+20.0%** | **SUPERIOR** |
72
+ | **Chemistry** | Science | 1/10 (10.0%) | **2/10 (20.0%)** | **+10.0%** | **SUPERIOR** |
73
+ | **Basic Medical Science** | Health & Medicine | 2/10 (20.0%) | **5/10 (50.0%)** | **+30.0%** | **SUPERIOR** |
74
+ | **Economics** | Business | 2/10 (20.0%) | **5/10 (50.0%)** | **+30.0%** | **SUPERIOR** |
75
+ | **Psychology** | Humanities & Social | 3/10 (30.0%) | 2/10 (20.0%) | -10.0% | Inferior |
76
+ | **Art Theory** | Art & Design | 7/10 (70.0%) | 5/10 (50.0%) | -20.0% | Inferior |
77
+ | **TOTAL GERAL (100 itens)** | **6 Áreas do Conhecimento** | **20/100 (20.0%)** | **33/100 (33.0%)** | **+13.0 p.p.** | **V12 SUPERIOR** |
78
+
79
+ > **Nota Metodológica Obrigatória:** A bateria acima constitui uma amostra experimental controlada de 100 questões do split de validação do MMMU padrão e **não equivale nem substitui a pontuação do benchmark completo ou do MMMU-Pro** (cujo valor oficial publicado pelo Google para o Gemma 4 E2B completo é de 44,2%).
80
+
81
+ ### Análise Estatística e Qualitativa:
82
+ - **Teste de Contingência McNemar:** $\chi^2 = 5,7600$ ($p = 0,0164$, ganho estatisticamente significativo com $p < 0,05$).
83
+ - **Redução Drástica de Evasão:** O modelo original gerou **52 respostas não interpretáveis (UNPARSED)** por esgotamento de tokens em dissertações reflexivas, enquanto o SpaceCode V12 reduziu esse número para apenas **14 UNPARSED** (-73,08% de respostas inválidas), respondendo de forma concisa e estruturada.
84
+
85
+ ---
86
+
87
+ ## 🧪 4. Resultados da Avaliação Funcional Independente (TASK-0015)
88
+
89
+ O SpaceCode-E2B V12 demonstrou conformidade integral nos testes formais pós-treinamento:
90
+ 1. **Engenharia de Software (Python 3.12 / Concorrência):** Geração idiomática com `asyncio.TaskGroup`, type hints completos e tratamento gracioso de cancelamento sob exceções.
91
+ 2. **Conhecimento Conceitual (Teorema CAP):** Síntese precisa dos trade-offs de consistência, disponibilidade e particionamento em bancos de dados distribuídos.
92
+ 3. **Tutoria Acadêmica (Escopo Léxico e Closure em JS):** Explicação didática clara estruturada em torno de um contador funcional sem jargões desnecessários.
93
+ 4. **Extração e Formatação Estrita:** Extração precisa de logs em formato JSON puro (`status_code`, `endpoint`, `horario`) sem ruído antes ou depois.
94
+ 5. **Anti-Alucinação Crítica (Sonda T04):** Recusa categórica e determinística a APIs inexistentes (rejeitando o hook fictício `useLifecycle()` no React Native e instruindo o usuário sobre métodos reais do ciclo de vida e `useEffect`).
95
+
96
+ ---
97
+
98
+ ## ⚠️ 5. Limitações Conhecidas
99
+
100
+ - **Execução em GPU de 8 GB de VRAM:** A viabilidade e a estabilidade de inferência em GPUs de consumo de 8 GB de VRAM **ainda não foi formalmente validada pela Arquitetura**. Embora a estimativa matemática aponte viabilidade sob quantização 4-bit, a homologação oficial para essa classe de hardware ocorrerá em etapa posterior.
101
+ - **Domínio Multimodal:** O modelo especializou predominantemente a geração de código, tutoria e interfaces móveis/web. Domínios visuais altamente abstratos (como Teoria da Arte pura) preservam as respostas herdadas da base ou podem apresentar variações estilísticas.
102
 
103
+ ---
104
+
105
+ ## 💻 6. Como Usar (Inferência com Hugging Face Transformers)
106
+
107
+ ### Instalação das Dependências:
108
  ```bash
109
+ pip install -q transformers accelerate safetensors torch pillow torchvision
 
110
  ```
111
 
112
+ ### Exemplo em Python (Texto e Visão Multimodal):
113
  ```python
114
+ import torch
115
+ from transformers import AutoProcessor, Gemma4ForConditionalGeneration
116
+ from PIL import Image
117
 
118
  MODEL_ID = "SpaceGhost/SpaceCode-E2B"
119
 
120
+ # Carregar processador e modelo completo merged em bfloat16
121
  processor = AutoProcessor.from_pretrained(MODEL_ID)
122
+ model = Gemma4ForConditionalGeneration.from_pretrained(
123
  MODEL_ID,
124
+ torch_dtype=torch.bfloat16,
125
  device_map="auto"
126
  )
127
 
128
+ # 1. Prompt de Código / Engenharia de Software:
129
+ prompt_code = [
130
  {
131
  "role": "user",
132
+ "content": "Como implementar um hook customizado em React Native com TypeScript para gerenciar conexões WebSocket com reconexão automática e limpeza de memória?"
 
 
 
133
  }
134
  ]
135
 
136
+ inputs = processor(
137
+ text=processor.apply_chat_template(prompt_code, tokenize=False, add_generation_prompt=True),
138
+ return_tensors="pt"
139
+ ).to("cuda")
140
+
141
+ with torch.no_grad():
142
+ outputs = model.generate(**inputs, max_new_tokens=400, do_sample=False)
143
 
144
+ response = processor.tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
 
145
  print(response)
146
+
147
+ # 2. Prompt Multimodal (com Imagem de Interface UI):
148
+ # img = Image.open("tela_aplicativo.png").convert("RGB")
149
+ # msgs = [
150
+ # {
151
+ # "role": "user",
152
+ # "content": [
153
+ # {"type": "image"},
154
+ # {"type": "text", "text": "Analise esta tela móvel e forneça os componentes React Native correspondentes."}
155
+ # ]
156
+ # }
157
+ # ]
158
+ # inp_multi = processor(text=processor.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True), images=[img], return_tensors="pt").to("cuda")
159
+ # out_multi = model.generate(**inp_multi, max_new_tokens=300, do_sample=False)
160
+ # print(processor.tokenizer.decode(out_multi[0][inp_multi.input_ids.shape[1]:], skip_special_tokens=True))
161
+ ```
162
+
163
+ ### Execução Local com Ollama (GGUF):
164
+
165
+ O modelo conta com versões quantizadas oficiais GGUF disponibilizadas no diretório `gguf/`:
166
+ - `SpaceCode-E2B-V12-Q8_0.gguf` (4,61 GiB — Alta Precisão de Referência)
167
+ - `SpaceCode-E2B-V12-Q4_K_M.gguf` (3,18 GiB — Execução Otimizada para Recursos Limitados)
168
+
169
+ #### Como Criar e Rodar no Ollama:
170
+ ```bash
171
+ # 1. Baixar o arquivo Modelfile e o binário GGUF Q4_K_M
172
+ huggingface-cli download SpaceGhost/SpaceCode-E2B Modelfile --local-dir .
173
+ huggingface-cli download SpaceGhost/SpaceCode-E2B gguf/SpaceCode-E2B-V12-Q4_K_M.gguf --local-dir .
174
+
175
+ # 2. Criar o modelo localmente no Ollama
176
+ ollama create spacecode-e2b -f Modelfile
177
+
178
+ # 3. Executar o modelo no terminal
179
+ ollama run spacecode-e2b
180
  ```
181
 
182
+ > **Nota sobre Multimodalidade no Ollama:** O suporte a imagens em GGUF no ecossistema Ollama/llama.cpp para a arquitetura Gemma 4 depende da integração completa do projetor visual `mtmd/gemma4v`. Para inferência multimodal integral com visão ativa (texto + imagem), utilize a biblioteca oficial `transformers` com os pesos Safetensors bfloat16 conforme demonstrado acima.
183
+
184
  ---
185
 
186
+ ## 📜 7. Histórico de Versões e Governança
187
 
188
+ | Versão | Tag da Release | Formato | Status | Destaques |
189
+ | :---: | :---: | :---: | :---: | :--- |
190
+ | **V12** | `v12` | **Merged 16-bit Safetensors** | **Atual Oficial** | Merge integral da base Gemma 4 E2B com LoRA V12. +13 p.p. no MMMU, 1.536 tokens, anti-alucinação T04. |
191
+ | **V11** | `v11` | Adapter LoRA (`specialization-task-0011`) | Homologado Baseline | Baseline imutável preservada (SHA-256: `cb58fbb0...`). |
192
+ | **V2** | `v2` | Merged 16-bit / GGUF | Histórico | Release inicial baseada na ementa Android/React Native (88% acurácia). |
 
193
 
194
  ---
195
 
196
+ ## ⚖️ 8. Licença e Atribuições
197
+
198
+ Este modelo baseia-se no **Google Gemma 4 E2B** e é disponibilizado em estrita conformidade com os termos da [Gemma Terms of Use](https://ai.google.dev/gemma/terms).
199
 
200
+ *Gemma is a trademark of Google LLC.*
201
+ *Desenvolvido e mantido pela equipe de Engenharia de Machine Learning da **SpaceGhost**.*
chat_template.jinja CHANGED
@@ -1,3 +1,9 @@
 
 
 
 
 
 
1
  {%- macro format_parameters(properties, required, filter_keys=false) -%}
2
  {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
3
  {%- set ns = namespace(found_first=false) -%}
@@ -247,17 +253,12 @@
247
  {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
248
  {%- endfor -%}
249
  {%- elif function['arguments'] is none -%}
250
- {%- elif function['arguments'] is string -%}
251
- {#- Pre-serialized args (e.g. an OpenAI JSON string). We cannot JSON-parse
252
- portably in-template, so render non-fatally instead of erroring. Strip an
253
- outer {...} so it composes with the DSL braces rather than double-wrapping.
254
- Prefer passing arguments as a mapping for exact Gemma DSL. -#}
255
- {%- set argstr = function['arguments'] | trim -%}
256
- {%- if argstr[:1] == '{' and argstr[-1:] == '}' -%}
257
- {{- argstr[1:-1] -}}
258
- {%- else -%}
259
- {{- function['arguments'] -}}
260
- {%- endif -%}
261
  {%- endif -%}
262
  {{- '}<tool_call|>' -}}
263
  {%- endfor -%}
 
1
+ {#
2
+ Template: Google Gemma 4 Canonical Chat Template
3
+ Author: Google Gemma Engineering Team
4
+ Published: 2026-07-09
5
+ Context: Fixed tool-calling loops, turn closures, and thinking content-ordering.
6
+ #}
7
  {%- macro format_parameters(properties, required, filter_keys=false) -%}
8
  {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
9
  {%- set ns = namespace(found_first=false) -%}
 
253
  {{- key -}}:{{- format_argument(value, escape_keys=False) -}}
254
  {%- endfor -%}
255
  {%- elif function['arguments'] is none -%}
256
+ {%- else -%}
257
+ {{- raise_exception(
258
+ "chat_template: tool_calls[].function.arguments must be a "
259
+ "JSON object (mapping), not a string. Deserialize arguments "
260
+ "before passing to the template."
261
+ ) -}}
 
 
 
 
 
262
  {%- endif -%}
263
  {{- '}<tool_call|>' -}}
264
  {%- endfor -%}
config.json CHANGED
@@ -3,38 +3,22 @@
3
  "Gemma4ForConditionalGeneration"
4
  ],
5
  "audio_config": {
6
- "_name_or_path": "",
7
- "architectures": null,
8
  "attention_chunk_size": 12,
9
  "attention_context_left": 13,
10
  "attention_context_right": 0,
11
  "attention_invalid_logits_value": -1000000000.0,
12
  "attention_logit_cap": 50.0,
13
- "chunk_size_feed_forward": 0,
14
  "conv_kernel_size": 5,
15
- "torch_dtype": "bfloat16",
16
  "gradient_clipping": 10000000000.0,
17
  "hidden_act": "silu",
18
  "hidden_size": 1024,
19
- "id2label": {
20
- "0": "LABEL_0",
21
- "1": "LABEL_1"
22
- },
23
  "initializer_range": 0.02,
24
- "is_encoder_decoder": false,
25
- "label2id": {
26
- "LABEL_0": 0,
27
- "LABEL_1": 1
28
- },
29
  "model_type": "gemma4_audio",
30
  "num_attention_heads": 8,
31
  "num_hidden_layers": 12,
32
- "output_attentions": false,
33
- "output_hidden_states": false,
34
  "output_proj_dims": 1536,
35
- "problem_type": null,
36
  "residual_weight": 0.5,
37
- "return_dict": true,
38
  "rms_norm_eps": 1e-06,
39
  "subsampling_conv_channels": [
40
  128,
@@ -45,13 +29,14 @@
45
  "audio_token_id": 258881,
46
  "boa_token_id": 256000,
47
  "boi_token_id": 255999,
48
- "torch_dtype": "bfloat16",
49
  "eoa_token_id": 258883,
50
  "eoa_token_index": 258883,
51
  "eoi_token_id": 258882,
52
  "eos_token_id": 106,
53
  "image_token_id": 258880,
54
  "initializer_range": 0.02,
 
55
  "model_type": "gemma4",
56
  "pad_token_id": 0,
57
  "text_config": {
@@ -59,7 +44,7 @@
59
  "attention_dropout": 0.0,
60
  "attention_k_eq_v": false,
61
  "bos_token_id": 2,
62
- "torch_dtype": "bfloat16",
63
  "enable_moe_block": false,
64
  "eos_token_id": 1,
65
  "expert_intermediate_size": null,
@@ -161,44 +146,29 @@
161
  "vocab_size_per_layer_input": 262144
162
  },
163
  "tie_word_embeddings": true,
164
- "transformers_version": "5.17.0",
165
  "unsloth_fixed": true,
 
166
  "video_token_id": 258884,
167
  "vision_config": {
168
- "_name_or_path": "",
169
- "architectures": null,
170
  "attention_bias": false,
171
  "attention_dropout": 0.0,
172
- "chunk_size_feed_forward": 0,
173
  "default_output_length": 280,
174
- "torch_dtype": "bfloat16",
175
  "global_head_dim": 64,
176
  "head_dim": 64,
177
  "hidden_activation": "gelu_pytorch_tanh",
178
  "hidden_size": 768,
179
- "id2label": {
180
- "0": "LABEL_0",
181
- "1": "LABEL_1"
182
- },
183
  "initializer_range": 0.02,
184
  "intermediate_size": 3072,
185
- "is_encoder_decoder": false,
186
- "label2id": {
187
- "LABEL_0": 0,
188
- "LABEL_1": 1
189
- },
190
  "max_position_embeddings": 131072,
191
  "model_type": "gemma4_vision",
192
  "num_attention_heads": 12,
193
  "num_hidden_layers": 16,
194
  "num_key_value_heads": 12,
195
- "output_attentions": false,
196
- "output_hidden_states": false,
197
  "patch_size": 16,
198
  "pooling_kernel_size": 3,
199
  "position_embedding_size": 10240,
200
- "problem_type": null,
201
- "return_dict": true,
202
  "rms_norm_eps": 1e-06,
203
  "rope_parameters": {
204
  "rope_theta": 100.0,
 
3
  "Gemma4ForConditionalGeneration"
4
  ],
5
  "audio_config": {
 
 
6
  "attention_chunk_size": 12,
7
  "attention_context_left": 13,
8
  "attention_context_right": 0,
9
  "attention_invalid_logits_value": -1000000000.0,
10
  "attention_logit_cap": 50.0,
 
11
  "conv_kernel_size": 5,
12
+ "dtype": "bfloat16",
13
  "gradient_clipping": 10000000000.0,
14
  "hidden_act": "silu",
15
  "hidden_size": 1024,
 
 
 
 
16
  "initializer_range": 0.02,
 
 
 
 
 
17
  "model_type": "gemma4_audio",
18
  "num_attention_heads": 8,
19
  "num_hidden_layers": 12,
 
 
20
  "output_proj_dims": 1536,
 
21
  "residual_weight": 0.5,
 
22
  "rms_norm_eps": 1e-06,
23
  "subsampling_conv_channels": [
24
  128,
 
29
  "audio_token_id": 258881,
30
  "boa_token_id": 256000,
31
  "boi_token_id": 255999,
32
+ "dtype": "bfloat16",
33
  "eoa_token_id": 258883,
34
  "eoa_token_index": 258883,
35
  "eoi_token_id": 258882,
36
  "eos_token_id": 106,
37
  "image_token_id": 258880,
38
  "initializer_range": 0.02,
39
+ "model_name": "unsloth/gemma-4-E2B-it",
40
  "model_type": "gemma4",
41
  "pad_token_id": 0,
42
  "text_config": {
 
44
  "attention_dropout": 0.0,
45
  "attention_k_eq_v": false,
46
  "bos_token_id": 2,
47
+ "dtype": "bfloat16",
48
  "enable_moe_block": false,
49
  "eos_token_id": 1,
50
  "expert_intermediate_size": null,
 
146
  "vocab_size_per_layer_input": 262144
147
  },
148
  "tie_word_embeddings": true,
149
+ "transformers_version": "5.21.0",
150
  "unsloth_fixed": true,
151
+ "unsloth_version": "2026.10.11",
152
  "video_token_id": 258884,
153
  "vision_config": {
 
 
154
  "attention_bias": false,
155
  "attention_dropout": 0.0,
 
156
  "default_output_length": 280,
157
+ "dtype": "bfloat16",
158
  "global_head_dim": 64,
159
  "head_dim": 64,
160
  "hidden_activation": "gelu_pytorch_tanh",
161
  "hidden_size": 768,
 
 
 
 
162
  "initializer_range": 0.02,
163
  "intermediate_size": 3072,
 
 
 
 
 
164
  "max_position_embeddings": 131072,
165
  "model_type": "gemma4_vision",
166
  "num_attention_heads": 12,
167
  "num_hidden_layers": 16,
168
  "num_key_value_heads": 12,
 
 
169
  "patch_size": 16,
170
  "pooling_kernel_size": 3,
171
  "position_embedding_size": 10240,
 
 
172
  "rms_norm_eps": 1e-06,
173
  "rope_parameters": {
174
  "rope_theta": 100.0,
generation_config.json CHANGED
@@ -2,7 +2,6 @@
2
  "bos_token_id": 2,
3
  "do_sample": true,
4
  "eos_token_id": [
5
- 1,
6
  1,
7
  106,
8
  50
@@ -11,5 +10,5 @@
11
  "temperature": 1.0,
12
  "top_k": 64,
13
  "top_p": 0.95,
14
- "transformers_version": "5.17.0"
15
  }
 
2
  "bos_token_id": 2,
3
  "do_sample": true,
4
  "eos_token_id": [
 
5
  1,
6
  106,
7
  50
 
10
  "temperature": 1.0,
11
  "top_k": 64,
12
  "top_p": 0.95,
13
+ "transformers_version": "5.21.0"
14
  }
Q4_K_M.gguf → model-00001-of-00003.safetensors RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:4830fd780f7e0acbdb2da5667de75bc14d5bc208a0e01ab5a84b91df41a93f44
3
- size 3427880384
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:67a2f4a7761f373975f639a83a97b6044353f4c682f8605776582377adfe55ad
3
+ size 1422278032
mmproj-F16.gguf → model-00002-of-00003.safetensors RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:140be8d7849741f88c50757d529b84373ee8e27052cc2236855b537f4a8215fa
3
- size 985654080
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:882030dc73cb8418a55a7e0b58ab6c81677330c19bb94518ea0a770542798057
3
+ size 4987173642
model.safetensors → model-00003-of-00003.safetensors RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:fc23d5f9c28db872956af6cc3ad2f9c78b2703cd630ccba311597dfc982deb59
3
- size 10246621918
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec074aac9b6bca91363560f8e719c7476491b7db8660ec9a3535aea261d840d0
3
+ size 3800219068
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer.json CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b1d8bced9d66859cd2c3f4dcd8ab427197d4d46af3d9598b72af9fcf80b8392e
3
- size 32169781
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cc8d3a0ce36466ccc1278bf987df5f71db1719b9ca6b4118264f45cb627bfe0f
3
+ size 32169626
tokenizer_config.json CHANGED
@@ -20,8 +20,7 @@
20
  "is_local": false,
21
  "local_files_only": false,
22
  "mask_token": "<mask>",
23
- "max_length": 2048,
24
- "model_max_length": 131072,
25
  "model_specific_special_tokens": {
26
  "audio_token": "<|audio|>",
27
  "boa_token": "<|audio>",
@@ -42,10 +41,8 @@
42
  "str_token": "<|tool_response>",
43
  "think_token": "<|think|>"
44
  },
45
- "pad_to_multiple_of": null,
46
  "pad_token": "<pad>",
47
- "pad_token_type_id": 0,
48
- "padding_side": "right",
49
  "processor_class": "Gemma4Processor",
50
  "response_schema": {
51
  "properties": {
@@ -88,210 +85,58 @@
88
  "type": "object",
89
  "x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?P<content>(?:(?!\\<turn\\|\\>)(?!\\<\\|tool_response\\>).)+)?(?:\\<turn\\|\\>|\\<\\|tool_response\\>)?"
90
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
91
  "soc_token": "<|channel>",
92
  "sot_token": "<|turn>",
93
  "stc_token": "<|tool_call>",
94
  "std_token": "<|tool>",
95
  "str_token": "<|tool_response>",
96
- "stride": 0,
97
  "think_token": "<|think|>",
98
  "tokenizer_class": "GemmaTokenizer",
99
- "truncation_side": "right",
100
- "truncation_strategy": "longest_first",
101
- "unk_token": "<unk>",
102
- "added_tokens_decoder": {
103
- "0": {
104
- "content": "<pad>",
105
- "single_word": false,
106
- "lstrip": false,
107
- "rstrip": false,
108
- "normalized": false,
109
- "special": true
110
- },
111
- "1": {
112
- "content": "<eos>",
113
- "single_word": false,
114
- "lstrip": false,
115
- "rstrip": false,
116
- "normalized": false,
117
- "special": true
118
- },
119
- "2": {
120
- "content": "<bos>",
121
- "single_word": false,
122
- "lstrip": false,
123
- "rstrip": false,
124
- "normalized": false,
125
- "special": true
126
- },
127
- "3": {
128
- "content": "<unk>",
129
- "single_word": false,
130
- "lstrip": false,
131
- "rstrip": false,
132
- "normalized": false,
133
- "special": true
134
- },
135
- "4": {
136
- "content": "<mask>",
137
- "single_word": false,
138
- "lstrip": false,
139
- "rstrip": false,
140
- "normalized": false,
141
- "special": true
142
- },
143
- "46": {
144
- "content": "<|tool>",
145
- "single_word": false,
146
- "lstrip": false,
147
- "rstrip": false,
148
- "normalized": false,
149
- "special": true
150
- },
151
- "47": {
152
- "content": "<tool|>",
153
- "single_word": false,
154
- "lstrip": false,
155
- "rstrip": false,
156
- "normalized": false,
157
- "special": true
158
- },
159
- "48": {
160
- "content": "<|tool_call>",
161
- "single_word": false,
162
- "lstrip": false,
163
- "rstrip": false,
164
- "normalized": false,
165
- "special": true
166
- },
167
- "49": {
168
- "content": "<tool_call|>",
169
- "single_word": false,
170
- "lstrip": false,
171
- "rstrip": false,
172
- "normalized": false,
173
- "special": true
174
- },
175
- "50": {
176
- "content": "<|tool_response>",
177
- "single_word": false,
178
- "lstrip": false,
179
- "rstrip": false,
180
- "normalized": false,
181
- "special": true
182
- },
183
- "51": {
184
- "content": "<tool_response|>",
185
- "single_word": false,
186
- "lstrip": false,
187
- "rstrip": false,
188
- "normalized": false,
189
- "special": true
190
- },
191
- "52": {
192
- "content": "<|\"|>",
193
- "single_word": false,
194
- "lstrip": false,
195
- "rstrip": false,
196
- "normalized": false,
197
- "special": true
198
- },
199
- "98": {
200
- "content": "<|think|>",
201
- "single_word": false,
202
- "lstrip": false,
203
- "rstrip": false,
204
- "normalized": false,
205
- "special": true
206
- },
207
- "100": {
208
- "content": "<|channel>",
209
- "single_word": false,
210
- "lstrip": false,
211
- "rstrip": false,
212
- "normalized": false,
213
- "special": true
214
- },
215
- "101": {
216
- "content": "<channel|>",
217
- "single_word": false,
218
- "lstrip": false,
219
- "rstrip": false,
220
- "normalized": false,
221
- "special": true
222
- },
223
- "105": {
224
- "content": "<|turn>",
225
- "single_word": false,
226
- "lstrip": false,
227
- "rstrip": false,
228
- "normalized": false,
229
- "special": true
230
- },
231
- "106": {
232
- "content": "<turn|>",
233
- "single_word": false,
234
- "lstrip": false,
235
- "rstrip": false,
236
- "normalized": false,
237
- "special": true
238
- },
239
- "255999": {
240
- "content": "<|image>",
241
- "single_word": false,
242
- "lstrip": false,
243
- "rstrip": false,
244
- "normalized": false,
245
- "special": true
246
- },
247
- "256000": {
248
- "content": "<|audio>",
249
- "single_word": false,
250
- "lstrip": false,
251
- "rstrip": false,
252
- "normalized": false,
253
- "special": true
254
- },
255
- "258880": {
256
- "content": "<|image|>",
257
- "single_word": false,
258
- "lstrip": false,
259
- "rstrip": false,
260
- "normalized": false,
261
- "special": true
262
- },
263
- "258881": {
264
- "content": "<|audio|>",
265
- "single_word": false,
266
- "lstrip": false,
267
- "rstrip": false,
268
- "normalized": false,
269
- "special": true
270
- },
271
- "258882": {
272
- "content": "<image|>",
273
- "single_word": false,
274
- "lstrip": false,
275
- "rstrip": false,
276
- "normalized": false,
277
- "special": true
278
- },
279
- "258883": {
280
- "content": "<audio|>",
281
- "single_word": false,
282
- "lstrip": false,
283
- "rstrip": false,
284
- "normalized": false,
285
- "special": true
286
- },
287
- "258884": {
288
- "content": "<|video|>",
289
- "single_word": false,
290
- "lstrip": false,
291
- "rstrip": false,
292
- "normalized": false,
293
- "special": true
294
- }
295
- },
296
- "chat_template": "{%- macro format_parameters(properties, required, filter_keys=false) -%}\n {%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}\n {%- set ns = namespace(found_first=false) -%}\n {%- for key, value in properties | dictsort -%}\n {%- set add_comma = false -%}\n {%- if not filter_keys or key not in standard_keys -%}\n {%- if ns.found_first %},{% endif -%}\n {%- set ns.found_first = true -%}\n {{ key }}:{\n {%- if value['description'] -%}\n description:<|\"|>{{ value['description'] }}<|\"|>\n {%- set add_comma = true -%}\n {%- endif -%}\n {%- if value['type'] | upper == 'STRING' -%}\n {%- if value['enum'] -%}\n {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}\n enum:{{ format_argument(value['enum']) }}\n {%- endif -%}\n {%- elif value['type'] | upper == 'ARRAY' -%}\n {%- if value['items'] is mapping and value['items'] -%}\n {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}\n items:{\n {%- set ns_items = namespace(found_first=false) -%}\n {%- for item_key, item_value in value['items'] | dictsort -%}\n {%- if item_value is not none -%}\n {%- if ns_items.found_first %},{% endif -%}\n {%- set ns_items.found_first = true -%}\n {%- if item_key == 'properties' -%}\n properties:{\n {%- if item_value is mapping -%}\n {{- format_parameters(item_value, value['items']['required'] | default([])) -}}\n {%- endif -%}\n }\n {%- elif item_key == 'required' -%}\n required:[\n {%- for req_item in item_value -%}\n <|\"|>{{- req_item -}}<|\"|>\n {%- if not loop.last %},{% endif -%}\n {%- endfor -%}\n ]\n {%- elif item_key == 'type' -%}\n {%- if item_value is string -%}\n type:{{ format_argument(item_value | upper) }}\n {%- else -%}\n type:{{ format_argument(item_value | map('upper') | list) }}\n {%- endif -%}\n {%- else -%}\n {{ item_key }}:{{ format_argument(item_value) }}\n {%- endif -%}\n {%- endif -%}\n {%- endfor -%}\n }\n {%- endif -%}\n {%- endif -%}\n {%- if value['nullable'] %}\n {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}\n nullable:true\n {%- endif -%}\n {%- if value['type'] | upper == 'OBJECT' -%}\n {%- if value['properties'] is defined and value['properties'] is mapping -%}\n {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}\n properties:{\n {{- format_parameters(value['properties'], value['required'] | default([])) -}}\n }\n {%- elif value is mapping -%}\n {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}\n properties:{\n {{- format_parameters(value, value['required'] | default([]), filter_keys=true) -}}\n }\n {%- endif -%}\n {%- if value['required'] -%}\n {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}\n required:[\n {%- for item in value['required'] | default([]) -%}\n <|\"|>{{- item -}}<|\"|>\n {%- if not loop.last %},{% endif -%}\n {%- endfor -%}\n ]\n {%- endif -%}\n {%- endif -%}\n {%- if add_comma %},{%- else -%} {%- set add_comma = true -%} {% endif -%}\n type:<|\"|>{{ value['type'] | upper }}<|\"|>}\n {%- endif -%}\n {%- endfor -%}\n{%- endmacro -%}\n{%- macro format_function_declaration(tool_data) -%}\n declaration:{{- tool_data['function']['name'] -}}{description:<|\"|>{{- tool_data['function']['description'] -}}<|\"|>\n {%- set params = tool_data['function']['parameters'] -%}\n {%- if params -%}\n ,parameters:{\n {%- if params['properties'] -%}\n properties:{ {{- format_parameters(params['properties'], params['required']) -}} },\n {%- endif -%}\n {%- if params['required'] -%}\n required:[\n {%- for item in params['required'] -%}\n <|\"|>{{- item -}}<|\"|>\n {{- ',' if not loop.last -}}\n {%- endfor -%}\n ],\n {%- endif -%}\n {%- if params['type'] -%}\n type:<|\"|>{{- params['type'] | upper -}}<|\"|>}\n {%- endif -%}\n {%- endif -%}\n {%- if 'response' in tool_data['function'] -%}\n {%- set response_declaration = tool_data['function']['response'] -%}\n ,response:{\n {%- if response_declaration['description'] -%}\n description:<|\"|>{{- response_declaration['description'] -}}<|\"|>,\n {%- endif -%}\n {%- if response_declaration['type'] | upper == 'OBJECT' -%}\n type:<|\"|>{{- response_declaration['type'] | upper -}}<|\"|>}\n {%- endif -%}\n {%- endif -%}\n }\n{%- endmacro -%}\n{%- macro format_argument(argument, escape_keys=True) -%}\n {%- if argument is none -%}\n {{- 'null' -}}\n {%- elif argument is string -%}\n {{- '<|\"|>' + argument + '<|\"|>' -}}\n {%- elif argument is boolean -%}\n {{- 'true' if argument else 'false' -}}\n {%- elif argument is mapping -%}\n {{- '{' -}}\n {%- set ns = namespace(found_first=false) -%}\n {%- for key, value in argument | dictsort -%}\n {%- if ns.found_first %},{% endif -%}\n {%- set ns.found_first = true -%}\n {%- if escape_keys -%}\n {{- '<|\"|>' + key + '<|\"|>' -}}\n {%- else -%}\n {{- key -}}\n {%- endif -%}\n :{{- format_argument(value, escape_keys=escape_keys) -}}\n {%- endfor -%}\n {{- '}' -}}\n {%- elif argument is sequence -%}\n {{- '[' -}}\n {%- for item in argument -%}\n {{- format_argument(item, escape_keys=escape_keys) -}}\n {%- if not loop.last %},{% endif -%}\n {%- endfor -%}\n {{- ']' -}}\n {%- else -%}\n {{- argument -}}\n {%- endif -%}\n{%- endmacro -%}\n{%- macro strip_thinking(text) -%}\n {%- set ns = namespace(result='') -%}\n {%- for part in text.split('<channel|>') -%}\n {%- if '<|channel>' in part -%}\n {%- set ns.result = ns.result + part.split('<|channel>')[0] -%}\n {%- else -%}\n {%- set ns.result = ns.result + part -%}\n {%- endif -%}\n {%- endfor -%}\n {{- ns.result | trim -}}\n{%- endmacro -%}\n\n{%- macro format_tool_response_block(tool_name, response) -%}\n {{- '<|tool_response>' -}}\n {%- if response is mapping -%}\n {{- 'response:' + tool_name + '{' -}}\n {%- for key, value in response | dictsort -%}\n {{- key -}}:{{- format_argument(value, escape_keys=False) -}}\n {%- if not loop.last %},{% endif -%}\n {%- endfor -%}\n {{- '}' -}}\n {%- else -%}\n {{- 'response:' + tool_name + '{value:' + format_argument(response, escape_keys=False) + '}' -}}\n {%- endif -%}\n {{- '<tool_response|>' -}}\n{%- endmacro -%}\n\n{#- ===== SETUP ===== -#}\n{%- set ns = namespace(prev_message_type=None, prev_non_tool_role=None) -%}\n{%- set loop_messages = messages -%}\n{%- set enable_thinking = enable_thinking | default(false) -%}\n{%- set preserve_thinking = preserve_thinking | default(false) -%}\n{{- bos_token -}}\n{#- Handle System/Tool Definitions Block -#}\n{%- if enable_thinking or tools or (messages and messages[0]['role'] in ['system', 'developer']) -%}\n {{- '<|turn>system\\n' -}}\n {#- Inject Thinking token at the very top of the FIRST system turn -#}\n {%- if enable_thinking -%}\n {{- '<|think|>\\n' -}}\n {%- set ns.prev_message_type = 'think' -%}\n {%- endif -%}\n {%- if messages and messages[0]['role'] in ['system', 'developer'] -%}\n {%- if messages[0]['content'] is string -%}\n {{- messages[0]['content'] | trim -}}\n {%- elif messages[0]['content'] is sequence -%}\n {%- for item in messages[0]['content'] -%}\n {{- item['text'] | trim + ' '-}}\n {%- endfor -%}\n {%- endif -%}\n {%- set loop_messages = messages[1:] -%}\n {%- endif -%}\n {%- if tools -%}\n {%- for tool in tools %}\n {{- '<|tool>' -}}\n {{- format_function_declaration(tool) | trim -}}\n {{- '<tool|>' -}}\n {%- endfor %}\n {%- set ns.prev_message_type = 'tool' -%}\n {%- endif -%}\n {{- '<turn|>\\n' -}}\n{%- endif %}\n\n{#- Pre-scan: find last user message index for reasoning guard -#}\n{%- set ns_turn = namespace(last_user_idx=-1) -%}\n{%- for i in range(loop_messages | length) -%}\n {%- if loop_messages[i]['role'] == 'user' -%}\n {%- set ns_turn.last_user_idx = i -%}\n {%- endif -%}\n{%- endfor -%}\n\n{#- Loop through messages -#}\n{%- for message in loop_messages -%}\n {%- if message['role'] != 'tool' -%}\n {%- set ns.prev_message_type = None -%}\n {%- set role = 'model' if message['role'] == 'assistant' else message['role'] -%}\n {#- Detect continuation using tracked state — O(1) instead of O(n) backward scan -#}\n {%- set continue_same_model_turn = (role == 'model' and ns.prev_non_tool_role == 'assistant') -%}\n {%- if not continue_same_model_turn -%}\n {{- '<|turn>' + role + '\\n' }}\n {%- endif -%}\n\n {#- Render reasoning/reasoning_content as thinking channel -#}\n {%- set thinking_text = message.get('reasoning') or message.get('reasoning_content') -%}\n {%- set thinking_gate = (loop.index0 > ns_turn.last_user_idx) or (preserve_thinking and message.get('tool_calls')) -%}\n {%- if thinking_text and thinking_gate -%}\n {{- '<|channel>thought\\n' + thinking_text + '\\n<channel|>' -}}\n {%- endif -%}\n\n {%- if message.get('tool_calls') -%}\n {%- for tool_call in message.get('tool_calls') -%}\n {%- set function = tool_call['function'] -%}\n {{- '<|tool_call>call:' + function['name'] + '{' -}}\n {%- if function['arguments'] is mapping -%}\n {%- set ns_args = namespace(found_first=false) -%}\n {%- for key, value in function['arguments'] | dictsort -%}\n {%- if ns_args.found_first %},{% endif -%}\n {%- set ns_args.found_first = true -%}\n {{- key -}}:{{- format_argument(value, escape_keys=False) -}}\n {%- endfor -%}\n {%- elif function['arguments'] is none -%}\n {%- elif function['arguments'] is string -%}\n {#- Pre-serialized args (e.g. an OpenAI JSON string). We cannot JSON-parse\n portably in-template, so render non-fatally instead of erroring. Strip an\n outer {...} so it composes with the DSL braces rather than double-wrapping.\n Prefer passing arguments as a mapping for exact Gemma DSL. -#}\n {%- set argstr = function['arguments'] | trim -%}\n {%- if argstr[:1] == '{' and argstr[-1:] == '}' -%}\n {{- argstr[1:-1] -}}\n {%- else -%}\n {{- function['arguments'] -}}\n {%- endif -%}\n {%- endif -%}\n {{- '}<tool_call|>' -}}\n {%- endfor -%}\n {%- set ns.prev_message_type = 'tool_call' -%}\n {%- endif -%}\n\n {%- set ns_tr_out = namespace(flag=false) -%}\n {%- if message.get('tool_responses') -%}\n {#- Legacy: tool_responses embedded on the assistant message (Google/Gemma native) -#}\n {%- for tool_response in message.get('tool_responses') -%}\n {{- format_tool_response_block(tool_response['name'] | default('unknown', true), tool_response['response']) -}}\n {%- set ns_tr_out.flag = true -%}\n {%- set ns.prev_message_type = 'tool_response' -%}\n {%- endfor -%}\n {%- elif message.get('tool_calls') -%}\n {#- OpenAI Chat Completions: forward-scan consecutive role:tool messages -#}\n {%- set ns_tool_scan = namespace(stopped=false) -%}\n {%- for k in range(loop.index0 + 1, loop_messages | length) -%}\n {%- if ns_tool_scan.stopped -%}\n {%- elif loop_messages[k]['role'] != 'tool' -%}\n {%- set ns_tool_scan.stopped = true -%}\n {%- else -%}\n {%- set follow = loop_messages[k] -%}\n {#- Resolve tool_call_id to function name -#}\n {%- set ns_tname = namespace(name=follow.get('name') or 'unknown') -%}\n {%- for tc in message.get('tool_calls') -%}\n {%- if tc.get('id') == follow.get('tool_call_id') -%}\n {%- set ns_tname.name = tc['function']['name'] -%}\n {%- endif -%}\n {%- endfor -%}\n {#- Handle content as string or content-parts array -#}\n {%- set tool_body = follow.get('content') -%}\n {%- if tool_body is string -%}\n {{- format_tool_response_block(ns_tname.name, tool_body) -}}\n {%- elif tool_body is sequence and tool_body is not string -%}\n {%- set ns_txt = namespace(s='') -%}\n {%- for part in tool_body -%}\n {%- if part.get('type') == 'text' -%}\n {%- set ns_txt.s = ns_txt.s + (part.get('text') | default('')) -%}\n {%- endif -%}\n {%- endfor -%}\n {{- format_tool_response_block(ns_tname.name, ns_txt.s) -}}\n {%- for part in tool_body -%}\n {%- if part.get('type') in ['image', 'image_url'] -%}\n {{- '<|image|>' -}}\n {%- elif part.get('type') in ['audio', 'input_audio'] -%}\n {{- '<|audio|>' -}}\n {%- elif part.get('type') == 'video' -%}\n {{- '<|video|>' -}}\n {%- endif -%}\n {%- endfor -%}\n {%- else -%}\n {{- format_tool_response_block(ns_tname.name, tool_body) -}}\n {%- endif -%}\n {%- set ns_tr_out.flag = true -%}\n {%- set ns.prev_message_type = 'tool_response' -%}\n {%- endif -%}\n {%- endfor -%}\n {%- endif -%}\n\n {%- set captured_content -%}\n {%- if message.get('content') is string -%}\n {%- if role == 'model' -%}\n {{- strip_thinking(message['content']) -}}\n {%- else -%}\n {{- message['content'] | trim -}}\n {%- endif -%}\n {%- elif message.get('content') is sequence -%}\n {%- for item in message['content'] -%}\n {%- if item.get('type') == 'text' -%}\n {%- if role == 'model' -%}\n {{- strip_thinking(item['text']) -}}\n {%- else -%}\n {{- item['text'] | trim -}}\n {%- endif -%}\n {%- elif item.get('type') in ['image', 'image_url'] -%}\n {{- '<|image|>' -}}\n {%- elif item.get('type') in ['audio', 'input_audio'] -%}\n {{- '<|audio|>' -}}\n {%- elif item.get('type') == 'video' -%}\n {{- '<|video|>' -}}\n {%- endif -%}\n {%- endfor -%}\n {%- endif -%}\n {%- endset -%}\n\n {{- captured_content -}}\n {%- set has_content = captured_content | trim | length > 0 -%}\n\n {#- Forward-scan: find next non-tool message role for continuation detection -#}\n {%- set next_nt = namespace(role=None, found=false) -%}\n {%- for j in range(loop.index0 + 1, loop_messages | length) -%}\n {%- if not next_nt.found -%}\n {%- if loop_messages[j]['role'] != 'tool' -%}\n {%- set next_nt.role = loop_messages[j]['role'] -%}\n {%- set next_nt.found = true -%}\n {%- endif -%}\n {%- endif -%}\n {%- endfor -%}\n\n {%- set continues_into_next = (\n role == 'model'\n and next_nt.role == 'assistant'\n and (not message.get('tool_calls') or ns_tr_out.flag)\n ) -%}\n\n {%- if ns.prev_message_type == 'tool_call' and not ns_tr_out.flag -%}\n {{- '<|tool_response>' -}}\n {%- elif continues_into_next -%}\n {%- elif not (ns_tr_out.flag and not has_content and not next_nt.found) -%}\n {{- '<turn|>\\n' -}}\n {%- endif -%}\n\n {#- Track previous non-tool role for next iteration (avoids O(n) backward scan) -#}\n {%- set ns.prev_non_tool_role = message['role'] -%}\n {%- endif -%}\n{%- endfor -%}\n\n{%- if add_generation_prompt -%}\n {%- if ns.prev_message_type != 'tool_response' and ns.prev_message_type != 'tool_call' -%}\n {{- '<|turn>model\\n' -}}\n {%- elif ns.prev_message_type == 'tool_response' and enable_thinking -%}\n {{- '<|channel>thought\\n' -}}\n {%- endif -%}\n{%- endif -%}\n"
297
- }
 
20
  "is_local": false,
21
  "local_files_only": false,
22
  "mask_token": "<mask>",
23
+ "model_max_length": 1000000000000000019884624838656,
 
24
  "model_specific_special_tokens": {
25
  "audio_token": "<|audio|>",
26
  "boa_token": "<|audio>",
 
41
  "str_token": "<|tool_response>",
42
  "think_token": "<|think|>"
43
  },
 
44
  "pad_token": "<pad>",
45
+ "padding_side": "left",
 
46
  "processor_class": "Gemma4Processor",
47
  "response_schema": {
48
  "properties": {
 
85
  "type": "object",
86
  "x-regex": "(\\<\\|channel\\>thought\\n(?P<thinking>.*?)\\<channel\\|\\>)?(?P<tool_calls>\\<\\|tool_call\\>.*\\<tool_call\\|\\>)?(?P<content>(?:(?!\\<turn\\|\\>)(?!\\<\\|tool_response\\>).)+)?(?:\\<turn\\|\\>|\\<\\|tool_response\\>)?"
87
  },
88
+ "response_template": {
89
+ "defaults": {
90
+ "role": "assistant"
91
+ },
92
+ "fields": {
93
+ "content": {
94
+ "close": [
95
+ "<turn|>",
96
+ "<|tool_response>",
97
+ "<eos>"
98
+ ],
99
+ "content": "text"
100
+ },
101
+ "thinking": {
102
+ "close": "<channel|>",
103
+ "content": "text",
104
+ "open": "<|channel>thought\n"
105
+ },
106
+ "tool_calls": {
107
+ "close": "<tool_call|>",
108
+ "content": "json",
109
+ "content_args": {
110
+ "string_delims": [
111
+ [
112
+ "<|\"|>",
113
+ "<|\"|>"
114
+ ]
115
+ ],
116
+ "unquoted_keys": true
117
+ },
118
+ "open_pattern": "<\\|tool_call>call:(?P<name>\\w+)",
119
+ "repeats": true,
120
+ "transform": {
121
+ "function": {
122
+ "arguments": "{content}",
123
+ "name": "{name}"
124
+ },
125
+ "type": "function"
126
+ }
127
+ }
128
+ },
129
+ "start_anchor": [
130
+ "<|turn>model\n",
131
+ "<tool_response|>"
132
+ ]
133
+ },
134
  "soc_token": "<|channel>",
135
  "sot_token": "<|turn>",
136
  "stc_token": "<|tool_call>",
137
  "std_token": "<|tool>",
138
  "str_token": "<|tool_response>",
 
139
  "think_token": "<|think|>",
140
  "tokenizer_class": "GemmaTokenizer",
141
+ "unk_token": "<unk>"
142
+ }